Applying the ML Hausman approach to the demand of water for residential use: heterogeneity vs two-error specification